LTX-2.5 22B dev generating synchronized audio and video in one sampling pass per stage, on the shape of the ComfyUI LTX-2.5 templates: text or image to video in two stages (half resolution, 2x latent upscale, a short distilled refine), first/last frame and reference-sheet (Ingredients IC-LoRA) in one stage, and an optional 2x pixel re-render through the Pixel Spatial Upscaler IC-LoRA.
63Graph nodes
35Node classes
23Inputs
9Models
Tags: motionltx-2.5audio-videotext-to-videoimage-to-videofirst-last-framereference-sheetrunpod-serverlessvolume-backed
Inputs (23)#
The typed parameter surface callers bind when they request this workflow. Enum options and numeric bounds are the values the workflow document declares.
modeenumdefault t2vt2v and i2v run the ComfyUI LTX-2.5 two-stage templates (video_ltx2_5_t2v / video_ltx2_5_i2v): sample at half the output size, upscale the latent 2x, refine 3 steps. i2v writes start_image into both stages' latents. flf2v runs the one-stage first/last-frame template (video_ltx2_5_flf2v): start_image and end_image are guides at frame 0 and the last frame. ingredients (A.01.03) conditions one stage on reference_sheet through the LTX-2.5 Ingredients IC-LoRA.
options:
t2vi2vflf2vingredientspromptstringrequireddefault a wooden boat creaking against a dock at dawn, gulls calling in the distanceWhat should happen AND what it should sound like; one prompt drives both tracks. LTX-2.5 multishot is written into this prompt (Lightricks' prompting guide); there is no per-shot input. In ingredients mode use the model card's two parts: 'Reference sheet: <the panels> Generated video: <the action>'.
negative_promptstringdefault pc game, console game, video game, cartoon, childish, uglyWhat to avoid, in both tracks. The default is the LTX-2.5 t2v/i2v templates' own; the Ingredients model card suggests 'worst quality, inconsistent motion, blurry, jittery, distorted'.
shotsarraydefault []Native multishot (A.01.02): 2 to 4 shot descriptions, composed into the one chronological paragraph Lightricks' prompting guide prescribes and used as the prompt. Write the first as a full sentence and each later one as its framing ('a medium close-up of her face as she...'); reuse the same identifiers for anyone who recurs. Leave prompt at its default when you set shots; both together are refused.
shot_transitionenumdefault hard_cutHow later shots are joined: hard_cut 'A hard cut jumps to', match_cut 'A match cut transitions to', dissolve 'The shot dissolves into'.
options:
hard_cutmatch_cutdissolvestart_imageimagei2v: the still the clip starts from. flf2v: the first frame. Required in those two modes, ignored in the others.
end_imageimageflf2v only: the last frame. Required in that mode, ignored in the others.
image_strengthfloatdefault 0.7min 0.0max 1.0i2v: LTXVImgToVideoInplace strength in stage one; flf2v: both LTXVAddGuide strengths. 0.7 in all three places of the LTX-2.5 templates. i2v's stage-two re-anchor stays at the template's 1.0.
reference_sheetimageingredients only: one composite image with a clean panel per character (face close-up and turnaround), prop and location, on a black background with no text (Ingredients model card). It is padded with black to 768x448 and repeated into the static control video the IC-LoRA reads. Required in that mode, ignored in the others.
ingredients_strengthfloatdefault 1.0min 0.0max 2.0ingredients only: strength_model of the Ingredients IC-LoRA. 1.0 is the model card's recommendation (the weights ship pre-scaled); Lightricks' example workflow ships 1.3.
generate_audiobooleandefault trueWhen false the audio track is not decoded or muxed. The audio latent is still allocated and sampled: LTX generates the two together, and removing it would change the video.
resolutionenumdefault 1280x704Output frame size. Every option is a multiple of 64 so the two-stage modes' half-size first stage stays on LTX's 32 px latent grid. ingredients ignores it and renders the Ingredients IC-LoRA's single trained bucket, 768x448 (model card: other resolutions are out of distribution). upscale pixel_x2 doubles both edges of whichever size rendered.
options:
1280x704704x12801024x576768x4481920x1088lengthintegerdefault 121min 9max 193Frames, 8n+1 (LTX's latent stride). 121 is 5 s at 24 fps, the templates' default and the Ingredients recipe. In ingredients mode the reference video is max(length, 121) frames, because the model card requires at least 121.
auto_durationbooleandefault falseLet the LTX-2.5 duration head size the clip from the prompt (ComfyUI LTXVDurationPredictor, snapped to 8k+1 frames). length becomes the upper bound, so the cost estimate stays an upper bound. t2v, i2v and flf2v only.
fpsenumdefault 24Frame rate, fed to LTXVConditioning, the audio latent and the muxer so both tracks share one clock. 24 is the LTX-2.5 templates' and the Ingredients recipe's rate.
options:
2425stepsintegerdefault 30min 1max 40speed_mode default only: dev-model steps in stage one. 30 is LTX_2_4_PARAMS in Lightricks' ltx-pipelines, which LTX-2.5 inherits; no ComfyUI LTX-2.5 template runs the non-distilled path.
cfg_videofloatdefault 3.0min 0.0max 20.0speed_mode default only: guidance on the video half in stage one (ltx-pipelines LTX-2.4+ video cfg_scale 3.0).
cfg_audiofloatdefault 7.0min 0.0max 20.0speed_mode default only: guidance on the audio half in stage one (ltx-pipelines audio cfg_scale 7.0, also LTXVDualCFGGuider's own default).
seedintegerdefault 0min 0Noise seed of the first stage. The refine stage and the pixel_x2 pass use the fixed seed 42 their templates ship.
samplerenumdefault eulerspeed_mode default only: the stage-one sampler.
options:
eulerdpmpp_2muni_pceuler_ancestralschedulerenumdefault simplespeed_mode default only: the stage-one schedule.
options:
simplenormalbetakarrasspeed_modeenumdefault defaultdistilled (V.12.03 parity) applies ltx-2.5-22b-distilled-lora-450 to the dev transformer at strength 1.0 and samples stage one on the 8-step distilled sigma list with CFG 1, the recipe of every LTX-2.5 template (which load the distilled transformer this volume does not carry); steps, cfg_video, cfg_audio, sampler and scheduler are ignored. The sampler is euler_ancestral (t2v/i2v templates), euler_ancestral_cfg_pp (Lightricks' Ingredients example) or SamplerEulerAncestral with eta 0 (flf2v template). The two-stage refine and the pixel_x2 pass always run distilled, in both modes.
options:
defaultdistilledupscaleenumdefault nonepixel_x2 (A.01.03) re-renders the finished clip at twice its width and height through the LTX-2.5 Pixel Spatial Upscaler IC-LoRA at strength 1.0, with the decoded clip as the half-resolution reference (the LoRA's reference_downscale_factor 2) and the first pass's audio frozen, on an 8-step distilled pass. Lightricks recommends a low base (~280p) for it; the per-job pixel cap refuses sizes whose doubled frame does not fit.
options:
nonepixel_x2ComfyUI node graph (63)#
The executable ComfyUI prompt graph: 63 nodes across 35 distinct node classes, wired by 85 data dependencies. Nodes tinted green come from a custom node pack this workflow declares; the rest are ComfyUI core / baked-community classes.
flowchart TD
n0["1: UNETLoader"]
n1["2: CLIPLoader"]
n2["3: VAELoader"]
n3["4: VAELoader"]
n4["5: LatentUpscaleModelLoader"]
n5["6: LoraLoaderModelOnly"]
n6["7: CLIPTextEncode"]
n7["8: CLIPTextEncode"]
n8["9: LTXVConditioning"]
n9["10: EmptyLTXVLatentVideo"]
n10["11: LoadImage"]
n11["12: ResizeImageMaskNode"]
n12["13: LTXVPreprocess"]
n13["14: LTXVImgToVideoInplace"]
n14["15: LoadImage"]
n15["16: ResizeImageMaskNode"]
n16["17: ResizeImageMaskNode"]
n17["18: LTXVPreprocess"]
n18["19: LTXVPreprocess"]
n19["20: LTXVAddGuide"]
n20["21: LTXVAddGuide"]
n21["22: LoadImage"]
n22["23: ResizeAndPadImage"]
n23["24: RepeatImageBatch"]
n24["25: LTXICLoRALoaderModelOnly"]
n25["26: LTXAddVideoICLoRAGuide"]
n26["27: LTXVEmptyLatentAudio"]
n27["28: LTXVConcatAVLatent"]
n28["29: LTXVDualCFGGuider"]
n29["30: KSamplerSelect"]
n30["31: SamplerEulerAncestral"]
n31["32: BasicScheduler"]
n32["33: ManualSigmas"]
n33["34: RandomNoise"]
n34["35: SamplerCustomAdvanced"]
n35["36: LTXVSeparateAVLatent"]
n36["37: LTXVCropGuides"]
n37["38: LTXVLatentUpsampler"]
n38["39: LTXVImgToVideoInplace"]
n39["40: LTXVConcatAVLatent"]
n40["41: RandomNoise"]
n41["42: LTXVDualCFGGuider"]
n42["43: KSamplerSelect"]
n43["44: ManualSigmas"]
n44["45: SamplerCustomAdvanced"]
n45["46: LTXVSeparateAVLatent"]
n46["47: VAEDecodeTiled"]
n47["48: LTXVAudioVAEDecode"]
n48["49: LTXICLoRALoaderModelOnly"]
n49["50: EmptyLTXVLatentVideo"]
n50["51: LTXAddVideoICLoRAGuide"]
n51["52: LTXVSetAudioRefTokens"]
n52["53: LTXVConcatAVLatent"]
n53["54: RandomNoise"]
n54["55: CFGGuider"]
n55["56: KSamplerSelect"]
n56["57: ManualSigmas"]
n57["58: SamplerCustomAdvanced"]
n58["59: LTXVSeparateAVLatent"]
n59["60: LTXVCropGuides"]
n60["61: VAEDecodeTiled"]
n61["62: CreateVideo"]
n62["63: SaveVideo"]
n0 -->|model| n5
n1 -->|clip| n6
n1 -->|clip| n7
n6 -->|positive| n8
n7 -->|negative| n8
n10 -->|input| n11
n11 -->|image| n12
n3 -->|vae| n13
n12 -->|image| n13
n9 -->|latent| n13
n10 -->|input| n15
n14 -->|input| n16
n15 -->|image| n17
n16 -->|image| n18
n8 -->|positive| n19
n8 -->|negative| n19
n3 -->|vae| n19
n9 -->|latent| n19
n17 -->|image| n19
n19 -->|positive| n20
n19 -->|negative| n20
n3 -->|vae| n20
n19 -->|latent| n20
n18 -->|image| n20
n21 -->|image| n22
n22 -->|image| n23
n8 -->|positive| n25
n8 -->|negative| n25
n3 -->|vae| n25
n9 -->|latent| n25
n23 -->|image| n25
n24 -->|latent_downscale_factor| n25
n2 -->|audio_vae| n26
n26 -->|audio_latent| n27
n0 -->|model| n31
n33 -->|noise| n34
n28 -->|guider| n34
n27 -->|latent_image| n34
n34 -->|av_latent| n35
n35 -->|latent| n36
n35 -->|samples| n37
n4 -->|upscale_model| n37
n3 -->|vae| n37
n3 -->|vae| n38
n12 -->|image| n38
n37 -->|latent| n38
n35 -->|audio_latent| n39
n5 -->|model| n41
n8 -->|positive| n41
n8 -->|negative| n41
n40 -->|noise| n44
n41 -->|guider| n44
n42 -->|sampler| n44
n43 -->|sigmas| n44
n39 -->|latent_image| n44
n44 -->|av_latent| n45
n3 -->|vae| n46
n2 -->|audio_vae| n47
n5 -->|model| n48
n8 -->|positive| n50
n8 -->|negative| n50
n3 -->|vae| n50
n49 -->|latent| n50
n46 -->|image| n50
n48 -->|latent_downscale_factor| n50
n8 -->|positive| n51
n8 -->|negative| n51
n50 -->|video_latent| n52
n51 -->|audio_latent| n52
n48 -->|model| n54
n50 -->|positive| n54
n50 -->|negative| n54
n53 -->|noise| n57
n54 -->|guider| n57
n55 -->|sampler| n57
n56 -->|sigmas| n57
n52 -->|latent_image| n57
n57 -->|av_latent| n58
n50 -->|positive| n59
n50 -->|negative| n59
n58 -->|latent| n59
n59 -->|samples| n60
n3 -->|vae| n60
n47 -->|audio| n61
n61 -->|video| n62
Nodes (63)#
1UNETLoadercoreunet_name = ltx-2.5-22b-dev-transformer-bf16.safetensorsweight_dtype = defaultoutputs:
MODEL2CLIPLoadercoreclip_name = gemma4-12b-with-proj-ltx-2.5-bf16.safetensorstype = ltxvdevice = defaultoutputs:
CLIP3VAELoadercorevae_name = ltx-2.5-audio-vae-bf16.safetensorsoutputs:
VAE4VAELoadercorevae_name = ltx-2.5-video-vae-bf16.safetensorsoutputs:
VAE5LatentUpscaleModelLoadercoremodel_name = ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensorsoutputs:
LATENT_UPSCALE_MODEL6LoraLoaderModelOnlycoremodel = ◂ node 1 · out[0]lora_name = ltx-2.5-22b-distilled-lora-450-bf16.safetensorsstrength_model = 1.0outputs:
MODEL7CLIPTextEncodecoreclip = ◂ node 2 · out[0]text = {{prompt}} tmploutputs:
CONDITIONING8CLIPTextEncodecoreclip = ◂ node 2 · out[0]text = {{negative_prompt}} tmploutputs:
CONDITIONING9LTXVConditioningcorepositive = ◂ node 7 · out[0]negative = ◂ node 8 · out[0]frame_rate = {{float(fps)}} tmploutputs:
CONDITIONINGCONDITIONING10EmptyLTXVLatentVideocorewidth = {{ingredients_size.width if mode == 'ingredients' else resolution_map[resolution].width // stage_one_divisor[mode]}} tmplheight = {{ingredients_size.height if mode == 'ingredients' else resolution_map[resolution].height // stage_one_divisor[mode]}} tmpllength = {{length}} tmplbatch_size = 1outputs:
LATENT11LoadImagecoreimage = {{start_image}} tmploutputs:
IMAGEMASK12ResizeImageMaskNodecoreinput = ◂ node 11 · out[0]resize_type = scale longer dimensionresize_type.longer_size = 1536scale_method = lanczosoutputs:
IMAGE13LTXVPreprocesscoreimage = ◂ node 12 · out[0]img_compression = 18outputs:
IMAGE14LTXVImgToVideoInplacecorevae = ◂ node 4 · out[0]image = ◂ node 13 · out[0]latent = ◂ node 10 · out[0]strength = {{image_strength}} tmplbypass = falseoutputs:
LATENT15LoadImagecoreimage = {{end_image}} tmploutputs:
IMAGEMASK16ResizeImageMaskNodecoreinput = ◂ node 11 · out[0]resize_type = scale dimensionsresize_type.width = {{resolution_map[resolution].width}} tmplresize_type.height = {{resolution_map[resolution].height}} tmplresize_type.crop = centerscale_method = nearest-exactoutputs:
IMAGE17ResizeImageMaskNodecoreinput = ◂ node 15 · out[0]resize_type = scale dimensionsresize_type.width = {{resolution_map[resolution].width}} tmplresize_type.height = {{resolution_map[resolution].height}} tmplresize_type.crop = centerscale_method = nearest-exactoutputs:
IMAGE18LTXVPreprocesscoreimage = ◂ node 16 · out[0]img_compression = 18outputs:
IMAGE19LTXVPreprocesscoreimage = ◂ node 17 · out[0]img_compression = 18outputs:
IMAGE20LTXVAddGuidecorepositive = ◂ node 9 · out[0]negative = ◂ node 9 · out[1]vae = ◂ node 4 · out[0]latent = ◂ node 10 · out[0]image = ◂ node 18 · out[0]frame_idx = 0strength = {{image_strength}} tmploutputs:
CONDITIONINGCONDITIONINGLATENT21LTXVAddGuidecorepositive = ◂ node 20 · out[0]negative = ◂ node 20 · out[1]vae = ◂ node 4 · out[0]latent = ◂ node 20 · out[2]image = ◂ node 19 · out[0]frame_idx = -1strength = {{image_strength}} tmploutputs:
CONDITIONINGCONDITIONINGLATENT22LoadImagecoreimage = {{reference_sheet}} tmploutputs:
IMAGEMASK23ResizeAndPadImagecoreimage = ◂ node 22 · out[0]target_width = {{ingredients_size.width}} tmpltarget_height = {{ingredients_size.height}} tmplpadding_color = blackinterpolation = lanczosoutputs:
IMAGE24RepeatImageBatchcoreimage = ◂ node 23 · out[0]amount = {{max(length, 121)}} tmploutputs:
IMAGE25LTXICLoRALoaderModelOnlycoremodel = {{model_source_map[speed_mode]}} tmpllora_name = ltx-2.5-22b-ic-lora-ingredients-0.9.safetensorsstrength_model = {{ingredients_strength}} tmploutputs:
MODELFLOAT26LTXAddVideoICLoRAGuidecorepositive = ◂ node 9 · out[0]negative = ◂ node 9 · out[1]vae = ◂ node 4 · out[0]latent = ◂ node 10 · out[0]image = ◂ node 24 · out[0]frame_idx = 0strength = 1.0latent_downscale_factor = ◂ node 25 · out[1]crop = disableduse_tiled_encode = falsetile_size = 256tile_overlap = 64outputs:
CONDITIONINGCONDITIONINGLATENT27LTXVEmptyLatentAudiocoreaudio_vae = ◂ node 3 · out[0]frames_number = {{length}} tmplframe_rate = {{float(fps)}} tmplbatch_size = 1outputs:
LATENT28LTXVConcatAVLatentcorevideo_latent = {{video_latent_source_map[mode]}} tmplaudio_latent = ◂ node 27 · out[0]outputs:
LATENT29LTXVDualCFGGuidercoremodel = {{model_source_map['ingredients' if mode == 'ingredients' else speed_mode]}} tmplpositive = {{positive_source_map[mode]}} tmplnegative = {{negative_source_map[mode]}} tmplvideo_cfg = {{1.0 if speed_mode == 'distilled' else cfg_video}} tmplaudio_cfg = {{1.0 if speed_mode == 'distilled' else cfg_audio}} tmploutputs:
GUIDER30KSamplerSelectcoresampler_name = {{distilled_sampler_map[mode] if speed_mode == 'distilled' else sampler}} tmploutputs:
SAMPLER31SamplerEulerAncestralcoreeta = 0.0s_noise = 1.0outputs:
SAMPLER32BasicSchedulercoremodel = ◂ node 1 · out[0]scheduler = {{scheduler}} tmplsteps = {{steps}} tmpldenoise = 1.0outputs:
SIGMAS33ManualSigmascoresigmas = 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0outputs:
SIGMAS34RandomNoisecorenoise_seed = {{seed}} tmploutputs:
NOISE35SamplerCustomAdvancedcorenoise = ◂ node 34 · out[0]guider = ◂ node 29 · out[0]sampler = {{sampler_source_map['eta0' if mode == 'flf2v' and speed_mode == 'distilled' else 'select']}} tmplsigmas = {{sigma_source_map[speed_mode]}} tmpllatent_image = ◂ node 28 · out[0]outputs:
LATENTLATENT36LTXVSeparateAVLatentcoreav_latent = ◂ node 35 · out[0]outputs:
LATENTLATENT37LTXVCropGuidescorepositive = {{positive_source_map[mode]}} tmplnegative = {{negative_source_map[mode]}} tmpllatent = ◂ node 36 · out[0]outputs:
CONDITIONINGCONDITIONINGLATENT38LTXVLatentUpsamplercoresamples = ◂ node 36 · out[0]upscale_model = ◂ node 5 · out[0]vae = ◂ node 4 · out[0]outputs:
LATENT39LTXVImgToVideoInplacecorevae = ◂ node 4 · out[0]image = ◂ node 13 · out[0]latent = ◂ node 38 · out[0]strength = 1.0bypass = falseoutputs:
LATENT40LTXVConcatAVLatentcorevideo_latent = {{stage_two_video_source_map[mode]}} tmplaudio_latent = ◂ node 36 · out[1]outputs:
LATENT41RandomNoisecorenoise_seed = 42outputs:
NOISE42LTXVDualCFGGuidercoremodel = ◂ node 6 · out[0]positive = ◂ node 9 · out[0]negative = ◂ node 9 · out[1]video_cfg = 1.0audio_cfg = 1.0outputs:
GUIDER43KSamplerSelectcoresampler_name = euler_ancestraloutputs:
SAMPLER44ManualSigmascoresigmas = 0.85, 0.7250, 0.4219, 0.0outputs:
SIGMAS45SamplerCustomAdvancedcorenoise = ◂ node 41 · out[0]guider = ◂ node 42 · out[0]sampler = ◂ node 43 · out[0]sigmas = ◂ node 44 · out[0]latent_image = ◂ node 40 · out[0]outputs:
LATENTLATENT46LTXVSeparateAVLatentcoreav_latent = ◂ node 45 · out[0]outputs:
LATENTLATENT47VAEDecodeTiledcoresamples = {{video_decode_source_map[mode]}} tmplvae = ◂ node 4 · out[0]tile_size = 512overlap = 64temporal_size = 64temporal_overlap = 16outputs:
IMAGE48LTXVAudioVAEDecodecoresamples = {{audio_latent_source_map[mode]}} tmplaudio_vae = ◂ node 3 · out[0]outputs:
AUDIO49LTXICLoRALoaderModelOnlycoremodel = ◂ node 6 · out[0]lora_name = ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensorsstrength_model = 1.0outputs:
MODELFLOAT50EmptyLTXVLatentVideocorewidth = {{2 * (ingredients_size.width if mode == 'ingredients' else resolution_map[resolution].width)}} tmplheight = {{2 * (ingredients_size.height if mode == 'ingredients' else resolution_map[resolution].height)}} tmpllength = {{length}} tmplbatch_size = 1outputs:
LATENT51LTXAddVideoICLoRAGuidecorepositive = ◂ node 9 · out[0]negative = ◂ node 9 · out[1]vae = ◂ node 4 · out[0]latent = ◂ node 50 · out[0]image = ◂ node 47 · out[0]frame_idx = 0strength = 1.0latent_downscale_factor = ◂ node 49 · out[1]crop = disableduse_tiled_encode = falsetile_size = 256tile_overlap = 64outputs:
CONDITIONINGCONDITIONINGLATENT52LTXVSetAudioRefTokenscorepositive = ◂ node 9 · out[0]negative = ◂ node 9 · out[1]audio_latent = {{audio_latent_source_map[mode]}} tmploutputs:
CONDITIONINGCONDITIONINGLATENT53LTXVConcatAVLatentcorevideo_latent = ◂ node 51 · out[2]audio_latent = ◂ node 52 · out[2]outputs:
LATENT54RandomNoisecorenoise_seed = 42outputs:
NOISE55CFGGuidercoremodel = ◂ node 49 · out[0]positive = ◂ node 51 · out[0]negative = ◂ node 51 · out[1]cfg = 1.0outputs:
GUIDER56KSamplerSelectcoresampler_name = euler_ancestraloutputs:
SAMPLER57ManualSigmascoresigmas = 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0outputs:
SIGMAS58SamplerCustomAdvancedcorenoise = ◂ node 54 · out[0]guider = ◂ node 55 · out[0]sampler = ◂ node 56 · out[0]sigmas = ◂ node 57 · out[0]latent_image = ◂ node 53 · out[0]outputs:
LATENTLATENT59LTXVSeparateAVLatentcoreav_latent = ◂ node 58 · out[0]outputs:
LATENTLATENT60LTXVCropGuidescorepositive = ◂ node 51 · out[0]negative = ◂ node 51 · out[1]latent = ◂ node 59 · out[0]outputs:
CONDITIONINGCONDITIONINGLATENT61VAEDecodeTiledcoresamples = ◂ node 60 · out[2]vae = ◂ node 4 · out[0]tile_size = 512overlap = 64temporal_size = 64temporal_overlap = 16outputs:
IMAGE62CreateVideocoreimages = {{final_images_map[upscale]}} tmplfps = {{float(fps)}} tmplaudio = ◂ node 48 · out[0]outputs:
VIDEO63SaveVideocorevideo = ◂ node 62 · out[0]filename_prefix = ltx25-avformat = mp4Parameter banks (16)#
The prompt / configuration lookup tables this workflow keys into from its inputs — the vocabulary that turns a style / palette / preset selection into graph parameters.
resolution_map (5)#
1280x704{"width": 1280, "height": 704}704x1280{"width": 704, "height": 1280}1024x576{"width": 1024, "height": 576}768x448{"width": 768, "height": 448}1920x1088{"width": 1920, "height": 1088}ingredients_size (2)#
width768height448stage_one_divisor (4)#
t2v2i2v2flf2v1ingredients1requires_families (1)#
ltx25model_source_map (3)#
default10distilled60ingredients250positive_source_map (4)#
t2v90i2v90flf2v210ingredients260negative_source_map (4)#
t2v91i2v91flf2v211ingredients261video_latent_source_map (4)#
t2v100i2v140flf2v212ingredients262stage_two_video_source_map (2)#
t2v380i2v390distilled_sampler_map (3)#
t2veuler_ancestrali2veuler_ancestralingredientseuler_ancestral_cfg_ppsampler_source_map (2)#
select300eta0310sigma_source_map (2)#
default320distilled330video_decode_source_map (4)#
t2v460i2v460flf2v372ingredients372audio_latent_source_map (4)#
t2v461i2v461flf2v361ingredients361final_images_map (2)#
none470pixel_x2610post_render (1)#
ltx25_shots_duration{"shots_input": "shots", "transition_input": "shot_transition", "prompt_input": "prompt", "mode_input": "mode", "auto_duration_input": "auto_duration", "length_input": "length", "fps_input": "fps", "duration_modes": ["t2v", "i2v", "flf2v"], "positive_text_node": "7", "conditioning_node": "9", "guider_node": "29", "duration_head_file": "ltx-2.5-duration-head-bf16.safetensors", "min_seconds": 1.0, "length_targets": [{"node": "10", "input": "length"}, {"node": "27", "input": "frames_number"}, {"node": "50", "input": "length"}]}Models & dependencies#
Custom node packs (1)#
The non-core ComfyUI node classes this graph requires; the RunPod worker image the workflow runs on must bake or install a pack that provides every one of them.
ComfyUI-LTXVideoModels required (9)#
ltx-2.5-22b-dev-transformer-bf16.safetensorsgemma4-12b-with-proj-ltx-2.5-bf16.safetensorsltx-2.5-video-vae-bf16.safetensorsltx-2.5-audio-vae-bf16.safetensorsltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensorsltx-2.5-22b-distilled-lora-450-bf16.safetensorsltx-2.5-22b-ic-lora-ingredients-0.9.safetensorsltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensorsltx-2.5-duration-head-bf16.safetensorsOutput contract#
What a successful run of this workflow returns.
primary{"type": "video", "format": "mp4", "codec": "h264", "fps_source": "declared", "audio": true, "alpha": false, "description": "Clip with generated audio muxed in."}Taxonomy & routing#
How the control plane classifies this workflow — from the committed workflow-taxonomy-registry.json. It drives the consistency / control surface the agentic director can exercise over the workflow.
assetFamilyjoint-av-sceneoutputPackageProfilevideo-master-profilecontrolModalitiesmodel-lockprompt-template-locksampler-scheduler-lockseed-lockkeyframe-anchor-locktemporal-lockreference-ensembleconsistencyDimensionsidentitymotionenvironmentnotesLTX-2.5 audio+video (A.01.02, A.01.03). Each stage samples ONE concatenated audio+video latent. keyframe-anchor-lock applies in i2v (LTXVImgToVideoInplace in both stages) and flf2v (LTXVAddGuide at the first and last frame); reference-ensemble and the identity dimension apply in ingredients mode, where one reference sheet of characters, props and location conditions the clip through the Ingredients IC-LoRA.